AI ROI: The Formula, a Worked Business Case and How to Measure It

TL;DR

• AI ROI is the net annual benefit of an AI investment divided by its full annual cost. The formula is simple. The work is in counting every cost line and only the benefits you can prove.
• Full cost means seats, tokens, integration and the people who run it. Leaving out tokens and people is the most common reason an AI business case looks better on paper than in the ledger.
• Count three kinds of benefit: time saved (discounted for how much of it turns into output), errors caught, and revenue protected. Take a baseline before rollout or you cannot measure any of them.
• Sentra lowers the token line by giving every model and agent the same compact, permission-aware company memory, and adds benefit by catching problems across ERP, CRM, HRIS and conversations before anyone asks.

What AI ROI means

AI ROI is the return a company gets from an AI investment, measured as net benefit divided by total cost over a set period, usually one year. The standard form is:

AI ROI (%) = (annual benefit minus annual cost) ÷ annual cost × 100

Two companion numbers make the ROI of AI easier to defend in an AI business case:

• Payback period (months) = first-year cost ÷ monthly benefit at steady state.
• Cost per outcome = annual cost ÷ number of outcomes (invoices checked, tickets resolved, briefs produced).

Measuring AI ROI goes wrong in predictable ways. Seat licences are counted but tokens are not. Hours saved are multiplied at full salary, as if every saved hour became output. Pilots launch without a baseline, so nothing can be compared afterwards. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value (gartner.com). Two of those four reasons are ROI problems.

The cost of AI: every line to include

The cost of AI has four parts. Put all four in the AI budget, even when a line sits in another department's budget today.

Seats

Seat prices are public for most assistants. Claude Team costs $20 per Standard seat per month billed annually, or $25 billed monthly, and Claude Enterprise lists at $20 per seat per month billed annually with a 20 seat minimum and usage billed separately at API rates (claude.com). Microsoft lists Microsoft Copilot (formerly Microsoft 365 Copilot) at $30 per user per month paid yearly, on top of a qualifying Microsoft 365 licence (microsoft.com). For a wider view, see our pages on Claude Enterprise and ChatGPT Enterprise pricing.

Tokens

Enterprise AI pricing is moving from seats to usage. Claude Enterprise bills usage at API rates on top of the seat. Microsoft prices Copilot Studio agents with Copilot Credits, sold in packs of 25,000 credits for $200 per month (microsoft.com). API prices vary by model. Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Opus 5.5 at $4 and $20 (claude.com).

Tokens are the line that grows with success. More users and more agent runs mean more tokens, and most of the input tokens in an agent run are context: the documents, records and history the model reads before it answers.

Integration

Integration covers connecting the systems the AI needs, mapping permissions, testing answers and maintaining connectors when source systems change. Price it as engineering weeks at a loaded rate, plus an annual upkeep share.

People

Every AI deployment needs an owner, someone to review output, and time for staff training. These costs are real even when no invoice arrives. Leave them out and the business case will not survive a finance review.

Benefit lines: what to count and how to discount it

Time saved. Measure it by timing a sample of real tasks before and after rollout. Do not rely on staff guesses of time saved. Then apply a realization rate, because a saved hour only counts if it turns into other work. The worked example below uses 50%. Test it after rollout.

Errors caught. Count only errors that were confirmed and either recovered or avoided: a duplicate invoice not paid, a wrong price corrected before billing. Unconfirmed flags are not benefit.

Revenue protected. This is the largest and least certain line. Count a renewal, order or price as protected only when the account owner confirms the AI flag changed the outcome, and agree an attribution rate in advance.

A worked AI business case with every assumption stated

The example below is illustrative. Replace each assumption with your own numbers. The company has 150 knowledge workers and is rolling out an assistant for everyone, plus one agent workflow in finance that checks invoices against purchase orders and contracts.

Assumptions

• 150 seats at $20 per seat per month, billed annually.
• The finance agent runs 10,000 times a month. Each run reads 150,000 input tokens and writes 5,000 output tokens, priced at $2 and $10 per million tokens.
• Integration takes two engineers six weeks, at a loaded cost of $200,000 per engineer per year.
• One owner spends 25% of their time on the program, at a loaded $180,000 per year.
• Training takes 2 hours per person at a loaded $75 per hour.
• Security and legal review costs $10,000.
• Time saved is 1 hour per person per week, over 46 working weeks, at $75 per hour, with a 50% realization rate.
• The finance agent finds 8 confirmed errors a month, averaging $1,200 each.
• Two renewals worth $40,000 a year each are saved after an AI flag, with 50% attribution.

First-year costs

First-year benefits

Result

• ROI: ($413,950 minus $201,654) ÷ $201,654 = 105% in year one.
• Payback: $201,654 ÷ ($413,950 ÷ 12) = 5.8 months, assuming benefits run at full rate from month one. Add the ramp-up months to be realistic.
• Year two: integration and most training fall away, so cost drops to about $123,000 for seats, tokens and the owner, plus any connector upkeep and refresher training. ROI rises for the same benefit.

Sensitivity

The realization rate moves the answer more than any other assumption. Show the board all three cases.

In this example tokens are $42,000, about a fifth of first-year cost and already the second largest line after integration. Double the agent runs, or move to a model that costs twice as much, and tokens become the largest line. Unlike integration, they recur every year.

Why token cost matters to AI ROI

Tokens are the only cost line that scales automatically with use. A seat costs the same whether someone asks five questions or five hundred. An agent that reads 150,000 tokens of context per run costs ten times as much in input tokens as one that reads 15,000, for the same answer.

Most of that context is padding. Agents re-read the same documents, pull whole threads to find one decision, and repeat background on every run because they have no memory between runs. Cutting what the model reads, without cutting what it knows, lowers the cost of AI directly.

Run your own figures through our agent token cost calculator, and see the best tools to reduce LLM token costs for options. The real token cost of AI agent memory explains how a memory layer changes the input token count.

How to measure AI ROI after rollout

1. Take the baseline before rollout. Time a sample of tasks, count current error rates, record cycle times. Keep the raw data.

2. Keep a control group where you can. One team or region without the tool for the first quarter gives you a fair comparison.

3. Track cost monthly, by line. Seats, tokens, integration upkeep and people time in one report. Our guide to AI agent observability covers how to see token use per agent.

4. Log confirmed benefits as they happen. Each error caught or renewal saved gets a date, an amount and a named person who confirmed it.

5. Re-run the business case each quarter. Replace assumptions with measured values. Cut use cases whose measured ROI stays below target after two quarters.

6. Report cost per outcome. Finance understands "$4 per invoice checked" faster than "105% ROI".

Where Sentra fits

Sentra affects both sides of the ROI equation without replacing your assistants or agents.

On the cost side, Sentra gives every model and agent the same company memory through REST and MCP, so ChatGPT, Claude, Copilot, Gemini and internal agents ask for the facts they need instead of re-reading raw documents on every run. On Terminal-Bench 2.1, agents using Sentra scored 88.31% against an 83.37% baseline, with 72.6% lower model cost and 41.2% fewer tokens.

On the benefit side, Sentra connects email, chat, meetings, documents, ERP, CRM, HRIS and BI, works out what is true now and who is working on what, and acts first. It finds what is going wrong, such as an invoice that contradicts a contract term agreed by email or a renewal at risk, works out who should act, and tells them. Those are the errors caught and revenue protected lines in the business case. A named person approves anything Sentra raises before it reaches anyone else, so each benefit has a confirmed owner, which is what makes it countable.

Sentra knows when a fact became true and when it stopped being true, so agents do not act on a superseded price or plan. Answers respect each source's existing permissions. It holds SOC 2 Type II and ISO 27001 and deploys in the cloud, in an isolated VPC, or air-gapped on premises.

At Lenskart, stalled dependencies across ten departments fell 58% in four weeks.

To build an AI business case that includes Sentra, book a demo at /demo.

How to choose which AI investments to fund

• Fund use cases with a measurable baseline first. If you cannot time it or count it today, you cannot prove ROI later.
• Prefer use cases that cross systems. The value of catching an invoice that contradicts a contract is higher than summarizing a single document, and harder for a person to do.
• Price tokens at 2x your pilot volume. Pilots run on small samples. Production runs on everything.
• Set an AI budget by line, not by tool. Seats, tokens, integration and people each need an owner. Tools bought team by team hide the total.
• Check readiness first. If ownership or data access is weak, ROI will be too. Our AI readiness assessment takes 15 minutes.

FAQ